Device and method for the generation of synthetic data in generative networks

Abstract

A method to generate synthetic data instances. The method includes generating a synthetic data instance for an input variable value of an input variable supplied to the generative model, classifying the synthetic data instance to generate a classification result, determining a loss function value of a loss function, the loss function evaluating the classification result and determining the gradient of the loss function with respect to the input variable. Depending on the absolute value of the gradient, the method includes generating a plurality of modified input variable values, determining, for each modified input variable value, the gradient of the loss function, combining the gradients of the loss function to generate an estimated gradient, and modifying the input variable value in a direction determined by the estimated gradient to generate a further input variable value. The generative model generates a further synthetic data instance for the further input variable value.

Description

CROSS REFERENCE[0001]The present application claims the benefit under 35 U.S.C. § 119 of European Patent Application No. EP 19205664.6 filed on Oct. 28, 2019, which is expressly incorporated herein by reference in its entirety.FIELD[0002]The present disclosure relates to a computer-implemented method for the generation of synthetic data instances using a generative model.BACKGROUND INFORMATION[0003]The research for deep generative models such as Generative Adversarial Nets (GANs) or Variational Autoencoders (VAE) has exploded over the last few years. These usually unsupervised models are showing very promising results, being nowadays already able to reproduce natural looking images at high resolution and at sufficient quality to even fool human observers.[0004]The future potential uses for deep generative models are considerable, ranging from anomaly detection, high quality synthetic image generation to providing explainability tools for datasets and models. This can for example ena...

AI Technical Summary

Benefits of technology

[0040]Such a device enables a more robust classification of input sensor data, and thus a more reliable control of the actuator.

[0041]For example, in the context of autonomous driving, a robust object classification and thus a reliable vehicle control can be achieved.

Problems solved by technology

However, these potential benefits come at the cost of a harder and more complex training compared to the standard classification models.

For example, during the training of a GAN, several problems may occur such as training instability, mode collapse, high variance gradients, vanishing generator gradients, etc.

Typically, the latent space of a generator is high-dimensional, so that an exhaustive search or analysis is intractable.

However, a problem may occur during the computations of the generator gradients where the gradients become very small or null.

This is a similar problem to the well-known “vanishing generator gradients” problem that can occur during the training of a GAN.

Thus, when the classification model used in the method is the discriminator of a GAN, the gradients of the loss function are likely to saturate, i.e., the gradients are likely to be null or smaller than the predefined threshold.

Images